Local masking in natural images measured via a new tree-structured forced-choice technique

Abstract

It is widely known that natural images can hide or mask visual signals, and that this masking ability can vary across different regions of the image. Previous studies have quantified masking by measuring image-wide detection thresholds or local thresholds for select image regions; however, little effort has focused on measuring local thresholds across entire images so as to achieve ground-truth masking maps. Such maps could prove invaluable for testing and refining masking models; however, obtaining these maps requires a prohibitive number of trials using a traditional forced-choice procedure. Here, we present a tree-structured forced-choice procedure (TS-3AFC) designed to efficiently measure local thresholds across images. TS-3AFC requires fewer trials than normal forced-choice by employing recursive patch subdivision in which the child patches are not tested individually until the target is detectable in the parent patch. We show that TS-3AFC can yield masking maps which demonstrate both intrasubject and inter-subject repeatability, and we analyze the performance of a modern masking model and two quality estimators in predicting the obtained ground-truth maps for a small set of images.